Speaker
Dr
Andreas Gellrich
(DESY)
Description
The vast majority of jobs in the Grid are embarrassingly parallel. In
particular HEP tasks are divided into atomic jobs without need for
communication between them. Jobs are still neither multi-threaded nor
multi-core capable. On the other hand, resource requirements reach
from CPU-dominated Monte Carlo jobs to network intense analysis jobs.
The main objective of any Grid site is to stably operate its Grid farm
while achieving a high job slot occupancy, an optimal usage of the
computing resources (network, CPU, memory, disk space) and guaranteed
shares for the VOs and groups. In order to optimize the utilization of
resources, jobs must be distributed intelligently over the slots,
CPUs, and hosts. Although the jobs resource requirements cannot be
deduced directly, jobs are mapped to POSIX user/group ID based on
their VOMS-proxy. The user/group ID allows to distinguish jobs,
assuming VOs make use of the VOMS group and role mechanism.
The multi-VO Tier-2 site at DESY (DESY-HH) supports ~20 VOs on
federated computing resources, using an opportunistic resource usage
model. As at many EGI/WLCG sites, the Grid farm is based on the
queuing system PBS/TORQUE, which was deployed from the EMI middleware
repositories. Initially, the scheduler MAUI was used. It showed severe
scalability problems with 4000 job slots as soon as the number of
running plus queued jobs approached 10000. Job scheduling became slow
or even blocked. In addition, MAUI's many configuration options
appeared to be hard to control.
To be able to further increase the number of worker nodes as
requested by the VOs (to currently 8000 job slots), DESY-HH needed a
scalable and performing scheduler, which runs in conjunction with
PBS/TORQUE. In the course of studying alternative scheduling models,
a home-made scheduler was developed (working title: MySched), which is
tailored to then needs of the DESY-HH Grid farm and uses the C-API of
PBS/TORQUE. It is based on a simple scheduling model without support
for multi-core jobs and job parallelism and is optimized for high
job slot occupancy and intelligent distribution of jobs to the worker
nodes. Furthermore, it allows for a fine-grained adjustment of
limits and parameters on VO and group level.
In the contribution to CHEP 2013 we will discuss the impact of a
classification of jobs according to their potential resources
requirements on scheduling strategies. Subsequently, we will describe
our home-made implementation and present operational results.
Author
Dr
Andreas Gellrich
(DESY)
